Fix more 16bit denorms in precision tests

Fix more 16bit denorms in precision tests

The denorm code that was fixed in: https://gerrit.khronos.org/#/c/4135/
still persisted in a different function that I hadn't noticed. Update it
there as well.

Because this function is being called on fixed floats, rather than on
random ones, I don't think the brokenness of the old code will cause a
practical problem, but the new version is simpler and more correct, so
I've changed it anyway.

Components: Vulkan
VK-GL-CTS issue: 1843
Affects: dEQP-VK.glsl.builtin.precision_fp16_storage32b.*

Change-Id: I6d6d7f8678f6edf41255e194650648c865b71e76